Monday, January 6, 2014
Nassim (The Regent Hotel)
Poster Board # P1-28

Brendan C. Delaney, MD1, Derek Corrigan, MSc2, Przemyslaw Kazienko, PhD3, Roxana Danger Mercaderes, Dr4, Tomasz Kajdanowicz, PhD3, Tomasz Wrobel3, Jean-Karl Soler, PhD5, Olga Kostopoulou, PhD1, Vasa Curcin, PhD4 and Thomas Fahey, PhD2, (1)King's College London, London, United Kingdom, (2)Royal College of Surgeons in Ireland, Dublin, Ireland, (3)Wroclaw University of Technology, Wroclaw, Poland, (4)Imperial College, London, United Kingdom, (5)Mediterranean Institute of Primary Care, Malta, Malta
Purpose: Diagnostic error accounts for the majority of litigation cases against Family Practitioners in Europe and the USA. Decision support for diagnosis, integrated with the electronic health record has been suggested as one part of dealing with this problem. Clinical Prediction Rules (CPRs), formal statements about the relationship between presenting problems, symptoms, signs and diagnoses are increasingly available. However, both a formal means of representing CPRs so that computer systems can use them, and a lack of evidence by which to populate them, have been barriers to adoption.

   Method: The TRANSFoRm project (, is a 5 year €9M EU project that is developing a digital infrastructure for the ‘learning healthcare system’ in Europe. Part of this project focuses on the delivery of a knowledge translation system for diagnosis in primary care. A formal representation for diagnostic evidence has been produced, along with a means of populating that evidence using both published rules and data mining.

   Result: We developed a formal ontology for diagnostic reasoning, with concepts for ‘reason for encounter’ (RFE), ‘diagnostic cue’, ‘prevalence’ and ‘diagnosis’. The ontology was developed in protégé and made available as a web-service using sesame. The ontology can support both numerical and categorical rules, along with the concept of ‘red flags’ (low predictive value cues for very serious outcomes). The ontology was populated with evidence from a review of literature for the diagnosis of chest pain, abdominal pain and shortness of breath in primary care. To further expand the knowledge available we used a data mining approach based on the KNIME tool. The TransHIS datset, consisting of 67,000 episodes of care with RFEs and diagnoses recorded by GPs in the Netherlands and Malta was used to derive classifiers, based on Naïve Bayes or Random Forest methods and then imported into the ontology.

   Conclusion: Using an ontology to store diagnostic information allows for three important requirements of an automated information system to support clinical diagnosis. 1. The system can provide a ranked list of potential diagnoses in real time as information is gathered. 2. The system can reason backwards to suggest relevant cues to gather based on potential diagnoses. 3. New data can be added, and the ontology will suggest new logical associations based on that data.